The Problem
Building production-grade LLM agents is fragmented: you juggle model APIs, tool calling, memory, multi-agent orchestration, and observability separately. Most frameworks force rigid orchestration patterns that fight modern model reasoning. Teams need a unified abstraction layer that works with model capabilities rather than constraining them.
What This Does
AgentScope is a Python-based agent framework (87 Python files) that provides production-ready abstractions: built-in ReAct agents, tools, skills, human-in-the-loop steering, memory, planning, realtime voice, evaluation, and model finetuning. The docs/tutorial/en/src/ directory shows the breadth—taskagent.py, taskmemory.py, taskrealtime.py, taskmcp.py, and workflowhandoffs.py cover the core feature set.
The framework emphasizes ecosystem integration. examples/agent/a2aagent/main.py demonstrates Agent-to-Agent protocol support, and examples/agent/a2uiagent/ includes a full TypeScript client (Lit-based) for interactive UI agents. The repo includes a Studio component (taskstudio.py) for visual tracing and steering of running agents. The examples/ directory (69 files) serves as the primary documentation, with patterns for multi-agent debate, routing, and concurrent workflows.
How To Use It
Setup: The package is published on PyPI (agentscope). Install with pip install agentscope. The pyproject.toml in examples/agent/a2uiagent/samples/generalagent/ confirms Python packaging.
Configuration: The framework requires model API keys, typically set as environment variables. The tutorial files (quickstartinstallation.py, quickstartagent.py) show the setup pattern. No central config file exists; configuration is code-based via the init function.
Running it: Start with the quickstart:
pip install agentscope Then run a tutorial example python docs/tutorial/en/src/quickstartagent.py
For the A2A example:
cd examples/agent/a2aagent python main.py
The TypeScript client in examples/agent/a2uiagent/samples/client/lit/ uses npm; package.json and vite.config.ts are present for local dev.
Real-World Use
A customer support system using A2A: an AgentScope agent handles initial triage via ReAct tool calling, then hands off to a specialized agent for billing or technical issues. The A2A protocol (a2aagent/agentcard.py) exposes the agent for discovery. For UI-heavy tasks, the A2UI agent (a2uiagent/samples/generalagent/) generates interactive forms—the UItemplete_examples/ folder shows booking, contact, and email compose forms—allowing the agent to render dynamic interfaces rather than static text.
Code Health & Issues
Low - Documentation-heavy repo: 117 doc files vs. 87 Python files. The tutorial covers 25+ topics, suggesting mature documentation but the actual framework code is compact—verify it meets your needs before committing. Low - Example-centric structure: Most Python files are in examples/, not a clear src/ package. The framework code is likely in the installed package, not this repo. This makes code review harder. Med - No visible test coverage: Only 5 test files for a framework of this scope. CI workflows exist (unittest.yml, pre-commit.yml) but test density is thin. Low - Recent feature churn: News entries show rapid additions (realtime voice, A2A, memory compression) across late 2025/early 2026. Expect API surface changes.
The Bottom Line
AgentScope is a serious, well-documented framework from a team with real investment (28k+ stars upstream). It's suited for teams building production agents who want built-in A2A, MCP, and observability rather than assembling them. The main risk is the example-heavy structure and thin test coverage—evaluate against your specific workflow before committing.